R-加快3维数组/矩阵创建速度的不同方法

Jay Mack 发表于 Dev

杰伊·麦克

我的问题是方法之一。我使用SO遍历方法在R中创建3维数组（这是我的第一个问题； R是约束）。用例是该最终数组需要经常更新，但是两个输入数组在不同的时间更新。目标是最大程度地减少最终阵列创建时间，而且还尽可能减少中间步骤。

我知道我可以使用Rcpp，而且为了提高可读性，我分配了比我所需更多的资源，但是我想知道的是：

是否有更好的方法来完成此操作？

if (!require("geosphere")) install.packages("geosphere")

#simulate real data
dimLength <- 418
latLong <- cbind(rep(40,418),rep(2,418))
potentialChurn <- as.matrix(rep(500,418))

#create 2D matrix
valueMat <- matrix(0,dimLength,dimLength)
value <- potentialChurn
valueTranspose <- t(value)
for (s in 1:dimLength){valueMat[s,] <- value + valueTranspose[s]}
diag(valueMat) <- 0

#create 3D matrix from copying 2D matrix
bigValMat <- array(0,dim=c(dimLength,dimLength,dimLength))
for (d in 1:dimLength){bigValMat[,d,] <- valueMat}

#get crow fly distance between locations, create 2D matrix
distMat <- as.matrix(outer(seq(dimLength), seq(dimLength), Vectorize(function(i, j) distCosine(latLong[i,], latLong [j,]))))

###create 3D matrix by calculating distance between any two locations;
# create 2D matrix from each column in original 2D matrix
# add this column-replicated 2D matrix to the original
bigDistMat <- array(0,dim=c(dimLength,dimLength,dimLength))
for (p in 1:dimLength){
  addCol <- distMat[,p]  
  addMatrix <- as.matrix(addCol)
  for (y in 2:dimLength) {addMatrix <- cbind(addMatrix,addCol)}
  bigDistMat[,p,] <- data.matrix(distMat) + data.matrix(addMatrix)}

#Final matrix calculation
bigValDistMat <- bigValMat / bigDistMat

...作为背景，这是针对使用巴塞罗那单车共享（Bicing）数据针对某类学生制定的提前两步预测政策的一部分。项目结束了，我对如何做得更好很感兴趣。

cmbarbu

通常，如果您想加快代码执行速度，则希望识别瓶颈并按照此处的说明进行修复。将您所有的代码放在一个函数中是一个好主意。

在您的特定情况下，R代码的for循环使用过多。您需要更多地向量化您的代码。

立即编辑以获得长答案：

    #simulate real data, you want them to be random
    dimLength <- 418
    latLong <- cbind(rnorm(dimLength,40,0.5),rnorm(dimLength,2,0.5))
    potentialChurn <- as.matrix(rnorm(dimLength,500,10))

    #create 2D matrix, outer is designed for this operation
    valueMat <- outer(value,t(value),FUN="+")[,1,1,]
    diag(valueMat) <- 0

    # create 3D matrix from copying 2D matrix, again, avoid for loop
    bigValMat <- array(rep(valueMat,dimLength),dim=c(dimLength,dimLength,dimLength))
    # and use aperm to permute the dimensions
    bigValMat <- aperm(bigValMat2,c(1,3,2))

    #get crow fly distance between locations, create 2D matrix
    # other packages are available to compute that kind of distance matrix
    # but let's stay in plain R
    # wordy but so much faster (and easier to read)
    longs1 <- rep(latLong[,1],dimLength)
    lats1 <- rep(latLong[,2],dimLength)
    latLong1 <- cbind(longs1,lats1)
    longs2 <- rep(latLong[,1],each=dimLength)
    lats2 <- rep(latLong[,2],each=dimLength)
    latLong2 <- cbind(longs2,lats2)
    distMat <- matrix(distCosine(latLong1,latLong2),ncol=dimLength)

    ###create 3D matrix by calculating distance between any two locations;
    # same logic than for bigValMat
    addMatrix <- array(rep(distMat,dimLength),dim=rep(dimLength,3))
    distMat3D <- aperm(addMatrix,c(1,3,2))
    bigDistMat <- addMatrix + distMat3D

    #get crow fly distance between locations, create 2D matrix
    #Final matrix calculation
    bigValDistMat <- bigValMat / bigDistMat